EmbedLLM: Learning Compact Representations of Large Language Models